Merge remote-tracking branch 'upstream/master'

README.md

@@ -150,7 +150,7 @@ DeepFaceLab is used by such popular youtube channels as
 
 <img src="doc/political_speech2.jpg" align="center">
 
-![](doc/youtube_icon.png) https://www.youtube.com/watch?v=sbFHhpYU15w
+![](doc/youtube_icon.png) https://www.youtube.com/watch?v=IvY-Abd2FfM
 
 <img src="doc/political_speech3.jpg" align="center">
 

core/leras/device.py

@@ -87,6 +87,7 @@ class Devices(object):
         os.environ['NN_DEVICES_INITIALIZED'] = '1'
         os.environ['NN_DEVICES_COUNT'] = '0'
         
+        os.environ['CUDA_​CACHE_​MAXSIZE'] = '2147483647'
         min_cc = int(os.environ.get("TF_MIN_REQ_CAP", 35))
         libnames = ('libcuda.so', 'libcuda.dylib', 'nvcuda.dll')
         for libname in libnames:

core/leras/nn.py

@@ -70,7 +70,6 @@ class nn():
                         first_run = True
                     os.environ['CUDA_CACHE_PATH'] = str(compute_cache_path)
 
-            #nvcuda.dll ignores this param : os.environ['CUDA_​CACHE_​MAXSIZE'] = '536870912' #512Mb (32mb default)
             os.environ['TF_MIN_GPU_MULTIPROCESSOR_COUNT'] = '2'
             os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3' # tf log errors only
 

core/leras/ops/__init__.py

@@ -342,10 +342,13 @@ def depth_to_space(x, size):
         x = tf.transpose(x, (0, 3, 4, 1, 5, 2))
         x = tf.reshape(x, (-1, oc, oh, ow))
         return x
-        
-        
 nn.depth_to_space = depth_to_space
 
+def pixel_norm(x, power = 1.0):
+    return x * power * tf.rsqrt(tf.reduce_mean(tf.square(x), axis=nn.conv2d_spatial_axes, keepdims=True) + 1e-06)
+nn.pixel_norm = pixel_norm
+
+
 def rgb_to_lab(srgb):
     srgb_pixels = tf.reshape(srgb, [-1, 3])
     linear_mask = tf.cast(srgb_pixels <= 0.04045, dtype=tf.float32)
@@ -376,6 +379,18 @@ def rgb_to_lab(srgb):
     return tf.reshape(lab_pixels, tf.shape(srgb))
 nn.rgb_to_lab = rgb_to_lab
 
+def total_variation_mse(images):
+    """
+    Same as generic total_variation, but MSE diff instead of MAE
+    """
+    pixel_dif1 = images[:, 1:, :, :] - images[:, :-1, :, :]
+    pixel_dif2 = images[:, :, 1:, :] - images[:, :, :-1, :]
+    
+    tot_var = ( tf.reduce_sum(tf.square(pixel_dif1), axis=[1,2,3]) +
+                tf.reduce_sum(tf.square(pixel_dif2), axis=[1,2,3]) )
+    return tot_var
+nn.total_variation_mse = total_variation_mse
+
 """
 def tf_suppress_lower_mean(t, eps=0.00001):
     if t.shape.ndims != 1: